Skills

The closest thing to this guide in installable form, and grouped the way the work actually divides: audience context, content and community, distribution and discovery, developer experience, growth, lifecycle. The design decision worth noticing is developer-audience-context — you state your persona, their technical level, their pain points and where they hang out once, and every other skill reads it first, which is what stops the outputs drifting into generic SaaS voice. Install that one even if you install nothing else; the folder list (hacker-news-strategy, docs-as-marketing, free-tier-strategy, developer-signup-flow) reads like a table of contents for the sections here.

The most-used shelf of the lot — 41.8k stars as of July 2026 — and the reason it earns a place on a developer-marketing page is structural rather than editorial. Every skill checks product-marketing first: one file at .agents/product-marketing.md holding your product, audience and positioning, so the claim is stated once and inherited by the landing page, the pricing page and the launch post instead of being re-improvised in each. That is the same discipline positioning demands of a human team, enforced by the tooling. The individual skills are generic B2B by default; the context file is where you make them survive contact with an engineer.

Where an auditor scores what you have, this decides what to write. Diátaxis’ four modes — tutorial, how-to, reference, explanation — expand into the fourteen types a developer product actually ships (quickstart, integration guide, migration guide, troubleshooting, SDK reference, config reference, runbook, glossary), each with its own guidance and anti-patterns. The pluggable style guides are the pragmatic part: a team picks Google or Stripe or minimal and inherits a settled answer on voice, person and tense, which is an argument you otherwise re-run in every review.

Docs are the highest-leverage surface you own and nobody scores their own. A thirty-three-check pass with a number at the end is worth having mostly because it converts “our docs are fine” into a claim someone can disagree with before a launch, and because the AI-discoverability half of the checklist is work most docs teams have not yet been asked to do. The same repo carries api-docs-quality-report for per-endpoint scoring and the linking audits (orphan-pages-internal-linking-opportunities, no-outlinks-audit) that catch the pages nothing points at — a common quiet failure in docs that grew by accretion.

For most developer products the README is the landing page: it is where a link from Hacker News lands, what a model retrieves when someone asks what your thing does, and the page read before anyone reaches the docs site. This is the only skill on the shelf that treats it as a scored artifact — every rule traces to a source (the canonical guides, twenty top-tier READMEs, GitHub’s own docs, and 2025–2026 research on how agents read repos), and review returns a rubric with concrete replacement text for the worst offender instead of vague prose. Reach for it before a launch, and again the moment an install command changes.

DevRel’s real input is what developers say when you are not in the room, and the reason most teams do it badly is that it is unglamorous, cross-platform, and never finished. Handing the sweep to an agent makes the cadence survivable: the same loop this paper’s scout desk runs on the industry, pointed at your category instead. Its siblings in the repo — reddit-engagement, hacker-news-strategy — cover what to do once you have heard something, including the part where posting in the wrong register gets you flagged.

For a growing share of your users the tool description is the documentation: they never read the reference, because an agent chose your tool on the strength of one sentence and a JSON schema. That makes naming, annotations (readOnlyHint, destructiveHint, idempotentHint) and error copy a first-touch surface, and it is on a marketing shelf because nobody else in the company is going to own how it reads. The phase most worth stealing is the last one: ten independent, verifiable evaluation questions that measure whether a model can complete real tasks against your server — a time-to-first-call test for the agent era.

Activation rarely dies at install; it dies in the hour after, when the first call returns an error nobody documented. The six sections this grades map almost exactly onto that hour, and the two teams usually miss — error handling and troubleshooting — are the two nobody is measured on. Run it on your own SDK docs and on the competitor a prospect is comparing you against; the delta is a more persuasive internal argument than a satisfaction survey.

A skill is a content surface that did not exist two years ago: your migration gotchas, your idiomatic error handling, your “do not do it that way” — installed into the tool a developer is already in, invoked without them ever visiting your site. This is the tool that builds and validates one, and the discipline it enforces is the discipline good docs need anyway: a description tight enough to trigger on the right task, progressive disclosure so the body stays small, no dead weight. The shelves catalogued on this page are themselves the evidence that the format travels — thirty-four skills, sixty skills, installed by strangers with one command.

Audit tools tell you what is missing; this changes how you brief the work in the first place. Twelve rules across four groups — content and strategy, technical implementation, authority and trust, platforms and analytics — each with templates and config you can lift. It is on the shelf because the rules are organised the way a marketer thinks rather than the way a crawler works, and because it ships in the format 37-plus agents can read, which is itself the argument its own category is making.

The third GEO entry here, and the one that goes past reporting: alongside brand and keyword research it carries improve-aeo-geo, which proposes framework-specific changes to the site itself. Pick between the three by what you are short of. Knowledge you lack: read the rules. A number you lack: run the llms.txt probe. A diff you lack, on a site you own and can review: this one. The marketing-engineer role it assumes — someone comfortable opening a pull request against the marketing site — is the role this whole shelf quietly requires.

The cheapest check on this page: minutes, no credentials, and it answers a question most devtool teams have not formally asked — whether the systems now fielding “what should I use for X” can reach and parse us at all. Treat the result as a floor rather than a strategy. A pass means you are legible; being chosen is a content problem, and blocking the crawlers you did not mean to block is a robots.txt problem you will be glad you found before the next quarter’s traffic review.

For most developer products the changelog is read more often than the blog and owned by nobody, because translating forty commits into user-facing language is a chore that loses to whatever shipped that week. Handing the first draft to an agent is the cheapest way to restore the cadence — and cadence is the point, as the operators who turned releases into a recurring weekly beat keep demonstrating. Keep the human edit: the categories come from the commits, but “you must change your webhook handler before August” never does.